JP5098504B2 - Character recognition program, character recognition device, and character recognition method - Google Patents

Description

  The present invention relates to a character recognition program, a character recognition device, and a character recognition method for recognizing characters in an input image, the character recognition program for recognizing a character image on which a pattern other than characters is superimposed, a character recognition device, and The present invention relates to a character recognition method.

  Conventionally, character patterns and their feature quantities are stored in advance as a dictionary for each character type in character recognition, a similarity is obtained by comparison with a recognition target image, and a character type with the highest similarity is output as a recognition result. Techniques have been used.

  In this way, when recognizing the character type from the feature amount of the character pattern, if the character is in contact with other characters or symbols in the recognition target image, the shape of the character pattern is damaged and the feature amount cannot be calculated correctly. , Recognition accuracy is significantly reduced. On the other hand, in the conventional technique, a technique for recognizing characters by separating a plurality of touched patterns is known.

  For example, Patent Document 1 regards a character string pattern composed of a plurality of character patterns in contact with each other as a cutting line of each character pattern from candidates of cutting line end points detected at the outer and inner contour portions of the character string pattern. By extracting a pair of cutting line end points that make up a line segment, and cutting out the character pattern at the line segment connecting the cutting line end points of this pair, it is possible to accurately cut out characters even in complicated and touching character strings Discloses a character recognition device.

  Patent Document 2 includes a character dictionary that does not contact a ruled line as a recognition dictionary and a character dictionary that contacts a ruled line as a recognition dictionary in character recognition for tabular books. There is disclosed a character recognition device that recognizes characters in a book with high accuracy by determining the presence or absence of contact between a character and a ruled line in an area and selecting a dictionary to be used for character recognition according to the determination result.

Japanese Patent Laid-Open No. 6-1111070 JP 2001-22889 A

  However, Patent Document 1 described above can be applied only when the character patterns are in contact with each other or when the specific shape and the character pattern are in contact like a circle character string. Similarly, the above-described Patent Document 2 can be applied only when a character pattern is in contact with a ruled line.

  That is, the above-described conventional technique has a problem in that character recognition cannot be sufficiently performed when the shape of the pattern contacting the character pattern is unknown. For this reason, for example, in order to mark a character string or number by handwriting directly in an application form or a choice field of a questionnaire and to read these contents by a computer, it is necessary to correctly recognize the character from a pattern in which the character and the mark overlap. However, since the mark made by the user has various shapes and various ways of contact with the character pattern, the conventional technology cannot sufficiently recognize the mark.

  Therefore, it has become an important issue to realize a technique for accurately recognizing a character image in which a pattern of an arbitrary shape is overlapped without depending on the shape of superimposed character patterns and marks.

  The present invention has been made to solve the above-described problems in the prior art and to solve the problems, and a character recognition program and a character recognition device for accurately recognizing character images with overlapping patterns of arbitrary shapes. An object of the present invention is to provide a character recognition method.

  In order to solve the above-described problems and achieve the object, the present invention divides an input image into a plurality of partial input images and recognizes a plurality of parts obtained by the division in character recognition for recognizing characters in the input image. A feature amount is calculated for each input image. Then, the feature amount of the partial character image in each part obtained by dividing the plurality of character images into a plurality of parts, the positional relationship of each partial character image in the character image, and the character type indicated by the character image, Are searched for partial character images having similar feature amounts for each of the partial input images. Further, a plurality of partial character images obtained by the search are totaled for each character type, and the positional relationship of the partial character images in each character type and the positional relationship in the input image of the partial input image corresponding to the partial character image The combination of partial input images that match the positional relationship of the partial character images is extracted as a character candidate. When the extracted character candidates have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as the constituent elements of the character indicated by the character type.

  Further, according to the present invention, in the above invention, among the plurality of partial input images, a partial input image similar to a partial character image of the same character type is used as a node, and the relative positional relationship between the partial character images corresponding to each node The method is characterized in that a graph is created by connecting nodes in which the relative positional relationship between the partial input images corresponding to each node does not contradict, and a clique of the created graph is extracted as a character candidate.

  Further, in the present invention, when the extracted character candidates have a predetermined number or more of partial input images, a rectangle circumscribing the combination of the partial input images is obtained and indicated by the character type in the rectangle. It is characterized by recognizing that a character exists.

  Further, according to the present invention, in the above invention, the search table is configured so that each partial character is obtained with respect to each partial character image obtained when division processing with different division sizes and / or division positions is performed on the same character image. The feature amount of the image, the positional relationship between the partial character images in the character image, and the character type indicated by the character image are stored in association with each other.

  Further, according to the present invention, in the above invention, the search table associates a plurality of similar feature amounts with the partial character image, and for each of the plurality of feature amounts, each of the partial character images in the character image. The positional relationship and the character type indicated by the character image are stored in association with each other.

  According to the present invention, the position of the character region and the character category can be obtained using the partial input image that does not overlap with the character pattern in the input image. There is an effect that characters can be recognized with high accuracy.

  In addition, according to the present invention, the positional relationship between a plurality of partial input images is graphed, and a combination of partial input images that is matched by clique extraction is obtained. The effect is that the characters can be recognized.

  In addition, according to the present invention, by making a rectangle circumscribing a combination of a plurality of matched partial input images as a character region, it is possible to easily specify a region where a character exists in the input image. Play.

  In addition, according to the present invention, it is possible to prevent a reduction in recognition accuracy due to a mesh division size or position shift with respect to an input image and to perform character recognition with high accuracy.

  In addition, according to the present invention, a partial character image similar to the partial input image can be easily searched by associating and registering a plurality of similar feature amounts with respect to the partial character image in the search table. There is an effect that the recognition speed can be improved.

  Embodiments of a character recognition program, a character recognition device, and a character recognition method according to the present invention will be described below in detail with reference to the drawings.

  In character recognition according to the present invention, a character pattern is recognized using a non-overlapping portion of a character pattern without separating a character pattern and a pattern other than characters from an input image. That is, as shown in FIG. 4, there is no mark even when the character string “1. Spouse” in the input image is marked by handwriting and it is difficult to separate the characters by connecting the characters. It is recognized from the feature of the part that the character “1. Spouse” exists.

  For this purpose, first, the input image is divided, and it is determined whether the character of each partial input image obtained by the division is similar to which part of which character. Then, when the positional relationship between a plurality of partial input images that are respectively similar to different portions of the same character is consistent with the positional relationship within the corresponding character, the plurality of partial input images are one of the characters. It is recognized as a part.

  FIG. 1 is a functional configuration diagram showing a functional configuration of character recognition according to an embodiment of the present invention. As shown in the figure, the character recognition device 1 includes an input unit 11, a display unit 12, an image reading unit 13, an external interface 14, a storage unit 15, and a control unit 20 therein.

  The input unit 11 is a processing unit that receives an operation input from an operator, and is realized by, for example, a keyboard. The display unit 12 is a device that performs display output to an operator, and is realized by, for example, a liquid crystal display.

  The image reading unit 13 is a processing unit that performs processing for reading an input image, and is realized by a scanner or the like. The external interface 14 is a processing unit that transmits and receives data by connecting to an external device.

  The storage unit 15 is a storage unit that stores various data used by the character recognition device 1 during processing and various data generated by the processing. In the example shown in FIG. A hash table 16 is stored. Here, the character category indicates a character type and a character name.

  The control unit 20 is a control unit that controls the entire character recognition apparatus 1 and includes a recognition processing unit 21 and a hash table registration unit 22 therein. The hash table registration unit 22 performs processing for creating the hash table 16 using the learning character image sample acquired through the external interface 14 and registering the hash table 16 in the storage unit 15.

  The recognition processing unit 21 is a processing unit that performs processing for recognizing characters from the input image read by the image reading unit 13. FIG. 2 is a functional configuration diagram illustrating the functional configuration of the recognition processing unit 21 in more detail. As shown in the figure, the recognition processing unit 21 includes a mesh division unit 31, a feature calculation unit 32, a normalization processing unit 33, a character category search unit 34, a position consistency determination unit 35, and a character determination unit 36 therein. Have

  The mesh dividing unit 31 performs a process of dividing the input image into a mesh and creating a partial input image. The feature calculation unit 32 performs a process of calculating the feature amount of the plurality of partial input images created by the mesh division unit 31. The normalization processing unit 33 performs processing for normalizing the feature amount calculated by the feature calculation unit 32.

  The character category search unit 34 refers to the hash table 16 using the feature amount normalized by the normalization processing unit 33 as a key, and searches for a partial character image of a character category having a similar feature amount for each partial input image.

  The position consistency determination unit 35 aggregates the plurality of partial character images obtained by the character category search unit 34 for each character category, and the positional relationship between the partial character images in each character category and the partial input image in the input image. A process of extracting a combination of partial input images that matches the positional relationship of the partial character images as a character candidate is performed.

  When the character candidates extracted by the position consistency determination unit 35 have a predetermined number or more of partial input images, the character determination unit 36 has a character category in which the partial input images of the character candidates are indicated by the character type. It is determined that it is a constituent element, and processing for displaying the character category on the display unit 12 is performed.

  Next, each process of the character recognition device 1 will be described in detail. First, the processing by the hash table registration unit 22 will be described with reference to FIG. FIG. 3 is an explanatory diagram for explaining the registration of the hash table 16 by the hash table registration unit 22.

  First, the hash table registration unit 22 acquires a character image sample for learning via the external interface 14, and divides the acquired character image into meshes by n × n (for example, n = 5). Then, each mesh obtained by the division is used as a partial character image of the character image, and the feature is calculated for each mesh (for each partial character image).

  Various methods can be applied to this feature calculation. For example, a weighted direction index histogram (Electronic Information and Communication Society paper D'87/7 Vol.J70-D No.7 pp.1390-1397'weighted direction index histogram) In this weighted direction index histogram, a vector having dimensions corresponding to the number of direction indices is obtained as a feature. In this embodiment, a four-dimensional index is obtained. The case where a feature vector is used will be described as an example.

  In the example illustrated in FIG. 3, the hash table registration unit 22 divides the character image of “arrangement”, which is a character sample for learning, by 5 × 5. Then, each mesh obtained by the division is regarded as a matrix of i rows and j columns, and each mesh is identified as “allocation” (1, 1) to “allocation” (5, 5) to calculate a feature vector of each mesh. , “Arrangement” (1, 1) has a feature vector of (29, 8, 13, 5), “Arrangement” (1, 2) has a feature vector of (32, 14, 18, 25), “Arrangement” (2 , 1) is (12, 2, 4, 37).

  Furthermore, when there are a plurality of learning character samples for the same character category, the characteristic components depending on the individual character images are removed by averaging the number of learning sample images belonging to the same character category, and the character Find the feature vector of the category itself.

  As a result, n × n mesh feature vectors are obtained for one character category. The calculation of the mesh feature vector is performed for each character category.

  Next, the hash table registration unit 22 converts the mesh feature vector into a hash value so that the character category and the mesh position can be subtracted from the hash value. The mesh feature vector is a vector having dimensions corresponding to the number of direction indexes, but normalization is performed so that the range is an integer from 0 to 9. As a result, the mesh feature vector takes values as powers (= 10000) of the number of direction indexes of 10 (= 4).

  Although any method can be used as the normalization method, it is preferable to adopt a method in which similar values are converted into the same value. For example, a method is preferable in which an integer quotient is obtained by dividing by a predetermined number, and when the quotient exceeds 9, it is forcibly replaced with 9.

  In the example illustrated in FIG. 3, the hash table registration unit 22 obtains an integer quotient by dividing each dimension value of the feature vector by “4”. As a result, the feature vector (29, 8, 13, 5) of “allocation” (1, 1) is normalized to (7, 2, 3, 3), and the feature vector of “allocation” (1, 2) ( 32, 14, 18, 25) is normalized to (8, 3, 4, 6), and the feature vector (12, 2, 4, 37) of “arrangement” (2, 1) is (3, 0, 1). , 9).

  Then, the hash table registration unit 22 registers the character category name and the mesh position (i, j) in the hash table 16 in association with the normalized mesh feature vector value. That is, when a mesh feature vector (va, vb, vc, vd) is given, normalization is performed on (va, vb, vc, vd) and converted to (Va, Vb, Vc, Vd), and H = Va × 1000 + Vb × 100 + Vc × 10 + Vd is calculated and (character category name, i, j) is recorded in H.

  In the example shown in FIG. 3, “allocation” (1, 1) whose normalized feature vector is (7, 2, 3, 3) is associated with the hash value (7233), and the normalized feature vector is “Arrangement” (1, 2) which is (8, 3, 4, 6) is associated with the hash value (8346), and “Arrangement” whose normalized feature vector is (3, 0, 1, 9) (2, 1) is associated with the hash value (3019).

  The hash table registration unit 22 creates the hash table 16 by performing all the above processes for the character category, and stores the hash table 16 in the storage unit 15. In the example shown in the figure, in the hash value (7233), “acid” (1,1), “king” (3,2) and the like are registered in association with “allocation” (1,1). In addition to “allocation” (2,1), “acid” (2,1), “ancestor” (1,3) and the like are registered in the hash value (3019) in association with each other, and the hash value (8346) is registered. In addition to “allocation” (1, 2), “ancestor” (3, 2), “酉” (1, 3) and the like are registered in association with each other.

  Next, processing by the recognition processing unit 21 will be described. FIG. 4 is an explanatory diagram illustrating input image feature calculation and character category search by the recognition processing unit 21. When an input image is input from the image reading unit 13, the mesh dividing unit 31 divides the input image into meshes as shown in FIG.

  At this time, the size of the mesh is set on the basis of the size by which one character in the input image is divided by n × n. For example, if the resolution of the input image is 400 dpi and the average character size is set to 40 pixels vertically and horizontally, and the mesh size is set to 8 pixels vertically and horizontally, one character is meshed in a size equivalent to 5 × 5. Can be divided. For other resolution images, the mesh size may be set in proportion to the resolution. If the size of the surrounding characters can be recognized, the size of the mesh may be set based on the size of the surrounding characters.

  Note that the mesh division unit 31 stores and stores in the storage unit 15 from which position of the input image each mesh (partial input image) obtained by mesh division of the input image is cut out.

  Next, the feature calculation unit 32 obtains a feature vector for each of the extracted meshes. For the calculation of the feature vector, for example, a weighted direction index histogram is used as in the case of creating the hash table. In the example shown in FIG. 4, the feature vector of the mesh m43 cut out from the input image is obtained as (13, 1, 5, 62), and the feature vector of the mesh m104 is obtained as (36, 7, 3, 4). ing.

  The normalization processing unit 33 performs processing for normalizing the feature vectors calculated by the feature calculation unit 32. This normalization is performed in the same manner as when creating the hash table. For example, an integer quotient is obtained by dividing by a predetermined number, and if the quotient exceeds 9, it is forcibly replaced with 9.

  In the example illustrated in FIG. 4, the normalization processing unit 33 divides the value of each dimension of the feature vector by “4” to obtain an integer quotient. As a result, the feature vector (13, 1, 5, 62) of the mesh m43 is normalized to (3, 0, 1, 9), and the feature vector (36, 7, 3, 4) of the mesh m104 is (9, 2,1,1).

  The character category search unit 34 refers to the hash table 16 using the normalized feature vector as a key, and searches for a partial character image of a character category having a similar feature amount for each of the partial input images.

  As a result, in the example shown in FIG. 4, as the partial character image similar to the mesh m43, the partial character image linked to the hash value (3019), that is, (2, 1) of the character category “arrangement”, the character category Search results include (2,1) of “acid”, (1,3) of character category “so”, and the like. Similarly, as the partial character image similar to the mesh m104, the partial character image linked to the hash value (9211), that is, (4, 4) of the character category “ancestor”, (5, 3) of the character category “acid” Etc. are obtained as search results.

  After the character category search unit 34 searches for all partial meshes cut out from the input image, that is, similar partial character images with respect to the partial input image, the position consistency determination unit 35 includes a plurality of partial characters obtained as search results. Aggregate images by character category.

  FIG. 5 is an explanatory diagram for explaining the tabulation for each character category of the search results. In the example shown in the figure, the search results of the meshes m43 and m104 are voted and added to the corresponding positions of the character category. Specifically, for the character category “arrangement”, the mesh m43 is voted at the position (2, 1), and the mesh m104 is voted at the position (5, 3). Similarly, the mesh m43 is voted for (2,1) of the character category “acid”, the mesh m43 is assigned to (1,3) of the character category “ances”, and the mesh m104 is assigned to (4,4) of the character category “ancestor”. Is voting.

  Next, the position consistency determination unit 35 compares the positional relationship in the input image with the positional relationship in the character category for the partial input image voted for each character category, and determines the consistency. Specifically, the position consistency determination unit 35 uses the partial input image voted for each position of the same character category as a node, and maintains the relationship between the meshes of the character category and the relationship between the meshes of the input image. A graph is generated by connecting nodes to each other with a path.

  FIG. 6 is an explanatory diagram for explaining the graph processing of the partial input image by the position consistency determination unit 35. In the example shown in the figure, mesh m21, mesh m43, mesh m44, mesh m104, mesh m105, and mesh m108, which are partial input images cut out from the input image, are all voted for the character category “arrangement”.

  Here, the mesh m21 is voted for (1, 1) of the character category “arrangement”. Similarly, the mesh m43 is (2,1), the mesh m44 is (2,2), the mesh m104 is (5,4), the mesh m105 is (5,5), and the mesh m108 is (4,4). ).

  The position consistency determination unit 35 uses all the partial input images (mesh m21, mesh m43, mesh m44, mesh m104, mesh m105, mesh m108) registered in the “arrangement” as nodes, and based on the positional relationship between the nodes. And draw a path.

  FIG. 7 is an explanatory diagram for explaining path connection between nodes. The node connection will be described with reference to the figure, taking as an example the positional relationship between the mesh m43, the mesh m105, and the mesh m108.

  First, consider the mesh m43 and the mesh m105. In the input image, the mesh m105 is positioned in the lower right direction with respect to the mesh m43. Also in the character category, the mesh m105 is located in the lower right direction with respect to the mesh m43. In this way, between the mesh m43 and the mesh m105, the relative positional relationship in the input image and the relative positional relationship in the character category are matched (the positional relationship is maintained). A path is connected between m43 and mesh m105 (see the graph shown in FIG. 6).

  Next, the mesh m105 and the mesh m108 are considered. In the input image, the mesh m108 is located in the right direction at the same height as the mesh m105. On the other hand, in the character category, the mesh m108 is positioned in the upper left direction with respect to the mesh m105. As described above, since the relative positional relationship in the input image and the relative positional relationship in the character category are not matched between the mesh m105 and the mesh m108, the mesh m105 and the mesh m108 have a gap between the mesh m105 and the mesh m108. Do not draw a path (see graph shown in FIG. 6).

  Further, consider the mesh m43 and the mesh m108. In the input image, the mesh m108 is positioned in the lower right direction with respect to the mesh m43. Also in the character category, the mesh m108 is located in the lower right direction with respect to the mesh m43. Thus, since the relative positional relationship in the input image and the relative positional relationship in the character category are matched between the mesh m43 and the mesh m108, the path between the mesh m43 and the mesh m108. (See the graph shown in FIG. 6).

  For each character category, the position consistency determination unit 35 checks whether or not the positional relationship is satisfied for all two relationships of the voted mesh, and creates a graph. Thereafter, the position consistency determination unit 35 performs a process of extracting a clique that is a complete subgraph from the created graph. The clique extracted by the position consistency determination unit 35 is a combination of partial input images that matches the positional relationship of the partial character images, and becomes a character candidate in the input image. Various algorithms can be used to extract a clique from a graph (for example, C. Bron and J. Kerbosch Algorithm 457. Finding all cliques of an undirected graph [H]. Comm. ACM. 16 (9) (September 1973).

  The character determination unit 36 evaluates the clique extracted by the position consistency determination unit 35 to identify where the character category matches in the recognition target area. Specifically, when the number of clique nodes is equal to or greater than a certain threshold, it is determined that the correspondence is correct, and it is determined that the character category exists in the target area corresponding to the node.

  FIG. 8 is an explanatory diagram for explaining clique extraction and evaluation. As shown in the figure, when a clique is extracted from the graph G1 having the mesh m21, the mesh m43, the mesh m44, the mesh m104, the mesh m105, and the mesh m108, the mesh m21, the mesh m43, the mesh m44, the mesh m104, and the mesh m105 are extracted. A clique G2 having five nodes and a clique G3 having four nodes of mesh m21, mesh m43, mesh m44, and mesh m108 are obtained.

  Here, if the determination threshold value by the character determination unit 36 is 5, the clique G2 has a number of nodes equal to or greater than the threshold value, so the character determination unit 36 determines that the area in the input image corresponding to each node of the clique G2 is a character. It is determined that the character area includes a character of category “Arrangement”. On the other hand, since the number of nodes of the clique G3 is less than the threshold value, the character determination unit 36 determines that the combination of the nodes of the clique G3 is not the “arrangement” character.

  More specifically, the character determination unit 36 obtains a rectangle circumscribing the partial input image corresponding to each node for the clique having the number of nodes equal to or greater than the threshold, and recognizes the circumscribed rectangle as a character region.

  FIG. 9 is an explanatory diagram for explaining recognition of a character area by the character determination unit 36. In the example shown in the figure, the character determination unit 36 determines the circumscribed rectangle of the mesh m21, mesh m43, mesh m44, mesh m104, and mesh m105 constituting the clique G2 as the character area A1, and the character area A1 includes “ Recognize that there is a character "composition".

  When there are a plurality of the same characters in the input image, the character determination unit 36 creates one graph from all the characters in the same category. Thereafter, a plurality of cliques whose number of nodes exceeds the threshold is extracted from the graph, and each clique forms a different character region in the input image.

  In the description so far, the operation of each processing unit has been described using conceptual drawings in order to clarify the positional relationship between the input image and the character category. Are stored in the storage unit 15 in a format suitable for the processing in FIG.

  For example, FIG. 10 shows voting result data created by the position consistency determination unit 35 from the search result obtained by the character category search unit. This voting result data is obtained as a result of counting by the position consistency determination unit 35 described with reference to FIG. 5, and is a table having three items: character category, position in category, and input image mesh. Retain data in the format.

  FIG. 11 is graph data created by the position consistency determination unit 35. In this graph data, for each node included in the graph, data is stored in the form of a table having a value “1” when a path is connected and a value “0” when a path is not connected. .

  Next, the flow of processing by the character recognition device 1 will be described. FIG. 12 is a flowchart for explaining the processing operation in hash table registration. In the process shown in the figure, first, the hash table registration unit 22 receives a plurality of sample character image data for each character category via the external interface 14 (step S101).

  Then, the hash table registration unit 22 divides the received sample character image data into meshes (step S102), and calculates a feature vector for each mesh obtained by the division, that is, each partial character image (step S103).

  Thereafter, the hash table registration unit 22 averages the feature vectors for each mesh position of the character category (step S104), and normalizes the averaged feature vectors (step S105). Then, the normalized feature vector is associated with the character category and the mesh position and registered in the hash table (step S106), and the process ends. The hash table registration unit 22 creates the hash table by executing the above processing for each character category.

  FIG. 13 is a flowchart for explaining the processing operation in the character recognition processing. In the process shown in the figure, first, the mesh dividing unit 31 divides the input image read by the image reading unit 13 into meshes (step S201). Next, the feature calculation unit 32 calculates a feature vector for each mesh (partial input image) obtained by the division (step S202). Then, the normalization processing unit 33 normalizes each feature vector (step S203), and the character category search unit 34 searches the hash table using the feature vector normalized for each mesh as a key (step S204).

  The position consistency determination unit 35 uses this search result to vote a feature vector for each character category (step S205), and displays a graph with nodes of input meshes (partial input images) voted for the same character category. Create (step S206). In this graph creation, as described above, the positional relationship between the input image of the partial image region corresponding to each node and the positional relationship within the character category are compared to draw a path between the nodes.

  Then, the position consistency determination unit 35 performs a process of extracting a clique from the graph of each character category (step S207). Of the obtained cliques, the character determination unit 36 determines that there is a character category corresponding to the area occupied by the clique constituent nodes for the clique whose number of nodes exceeds the threshold (step S208), and ends the process.

  Next, modified examples of hash table registration and character recognition will be described. When searching for partial character images whose features are similar to those of the partial input image, for example, the similarity is calculated from the distance between the feature vector of the partial character image and the feature vector of the partial character image. Although it is also possible to use a method for determining that, it takes time to calculate a similar combination by measuring a distance between vectors. Therefore, the character recognition device 1 speeds up the recognition process by converting the feature vector of the partial character image into a hash value and subtracting the character category and the mesh position from the hash value.

  Specifically, when creating a hash value from a feature vector, the character recognition device 1 simplifies the calculation of the similarity by normalizing so that similar values are converted to the same value. In the above-described example, a case has been described in which a method is used in which the value of each dimension of the feature vector is divided by a predetermined number to obtain an integer quotient, and when the quotient exceeds 9, a method of forcibly replacing with 9 is adopted. A modification of the technique will be described.

  FIG. 14 is an explanatory diagram for explaining a modification example of normalization when creating a hash table. In the example shown in the figure, an integer a and an integer b (a> b) are defined, and an integer quotient of (xi ± b) / a with respect to a four-dimensional vector value (x1, x2, x3, x4). Ask for. If this quotient exceeds 9, it is forcibly replaced with 9.

  For example, if the (2,1) feature vector of the character category “arrangement” is (12,2,4,37), and a = 4 and b = 1, the normalized value of x1 is (12 + 1) / 4. = 3 and (12-1) / 4 = 2, with x2 normalized (2 + 1) / 4 = 0 and (2-1) / 4 = 0, with x3 normalized (4 + 1) / 4 = 1 and (4-1) / 4 = 0, and (37 + 1) / 4 = 9 and (37-1) / 4 = 9 are obtained by normalizing x4. Then, when all combinations that can be obtained are obtained, (3,0,0,9), (3,0,1,9), (4,0,0,9), (4,0,1,9) Four combinations are obtained. In this case, four feature vectors corresponding to (2, 1) of the character category “arrangement” are registered in the hash table.

  In this way, a hash value to be associated with a certain feature vector has a width, and a plurality of hash values are associated with a combination of a character category and a mesh position and registered in a hash table. It is possible to obtain an effect similar to that obtained when the value is lowered, that is, an effect of acquiring a feature vector having a relatively low similarity as a search result.

  By the way, when mesh division is performed on a sample character image or an input image, the value of the mesh feature vector changes depending on the position of the mesh. Therefore, when creating a hash table, it is desirable to consider a mesh shifted to some extent and register mesh feature vectors of sample characters in multiple. For an input image, a mesh feature vector may be calculated using a single mesh. Similarly, regarding the size of the mesh, a plurality of sizes are considered, and mesh feature vectors of characters are registered in multiple.

  FIG. 15 is an explanatory diagram for explaining a case where a plurality of different mesh divisions are performed on one character image. In the example shown in the figure, when a character image is divided into n × n, it is divided into three mesh sizes of n = 4, 5, and 6. In addition to this, for each divided size, a feature is similarly obtained for a mesh shifted by several pixels in the x and y directions.

  A feature vector obtained by performing mesh division without deviation in the x direction and y direction with the division number n is a (n, 0, 0) feature of the character category, and a mesh having deviations in the x direction and y direction with the division number n A feature vector obtained by dividing is defined as an (n, dx, dy) feature. For example, dx and dy are (n, 0, 0) feature, (n, 0, 1) feature, (n, 0, 2) feature by setting two values to divide the mesh into three equal parts. , (N, 1,0) features, (n, 1,1) features, (n, 1,2) features, (n, 2,0) features, (n, 2,1) features, (n, 2 , 2) Nine features are set. For example, when one side of a mesh is 6 pixels, the mesh can be divided into 3 equal parts by shifting by 2 pixels.

  Thus, in this example, 27 (n, dx, dy) features (n = 4, 5, 6; dx = 0, 1, 2; dy = 0, 1, 2) are obtained and registered in the hash table, respectively. To do. The value of the feature is (n, dx, dy) − (i, j) − (v1) based on the number of rows i, the number of columns j and the feature vector (v1, v2, v3, v4) when the mesh is regarded as a matrix. , v2, v3, v4). Therefore, when registering in the hash table, (character category name, n, dx, dy, i, j) is registered in association with the hash value H calculated from the feature vector (v1, v2, v3, v4). That's fine.

  In this way, in the recognition processing when (n, dx, dy) features are obtained and registered in the hash table, a plurality of partial character images having different mesh sizes and mesh positions are obtained as search results. Each search result is projected to evaluate the consistency of the positional relationship.

  FIG. 16 is an explanatory diagram for explaining the recognition processing when the (n, dx, dy) feature is used. In the example shown in the figure, there are partial character images corresponding to the (4,0,0) feature and the (5,0,0) feature for the mesh mα on the input image. There is a partial character image corresponding to the mesh mβ with (4, 0, 1) feature. In this case, by projecting the position of each partial character image on the character category, mα ′ as the projected image at (4,0,0) and mα as the projected image at (5,0,0) for the mesh mα. '' Is obtained. Similarly, mβ ′ is obtained as a projection image at (4, 0, 1) for the mesh mβ.

  Thus, even if feature vectors having different mesh division sizes and positions are mixed, the mutual positional relationship can be evaluated by projecting onto the character category. As shown in the figure, when a plurality of projection images mα ′ and mα ″ are obtained from one partial input image mα, each projection image may be handled as an individual node.

  Next, an application example of the present invention will be described. FIG. 17 is a schematic configuration diagram illustrating a schematic configuration of an image recognition system using the present invention. An image recognition system 100 shown in FIG. 1 is connected to a scanner 101 and acquires image data D1 read by the scanner 101. The image data D1 is an application form or a questionnaire in which character strings and numbers are directly handwritten in the option field, and handwritten marks are duplicated in the character pattern.

  The image recognition system 100 includes a character recognition device 1, a difference image creation unit 102, and an image analysis unit 103 therein. As described so far, the character recognition device 1 performs character recognition on the image data D1 and outputs image data D2 as a recognition result. This image data D2 indicates what character is present at which position in the image.

  The difference image creation unit 102 creates the image data D3 by taking the difference between the image data D1 and the image data D2. Since the image data D1 has a handwritten mark superimposed on the character and the image data D2 includes only the character, the image data D3 which is the difference between them is an image from which the handwritten mark has been extracted.

  The image analysis unit 103 outputs which option is selected from the position of the handwritten mark shown in the image data D3, the character shown in the image data D2, and the position thereof. In the example shown in the figure, the image analysis unit 103 can output analysis result data D4 indicating that the image data D1 is selected from “1. Spouse”.

  As described above, in the character recognition device 1 according to the present embodiment, the hash table registration unit 22 creates the hash table 16 indicating the characteristics of the partial input image that is a local area of each character category, and performs recognition processing. The unit 21 meshes and divides the input image into partial input images, calculates features for each partial input image, retrieves a partial character image having features similar to each partial input image, and searches the partial input image Are compared with the positional relationship of the partial character images, and the consistency is evaluated to recognize which character is present in which region of the input image.

  Therefore, the character recognition device 1 can recognize a character by using a non-overlapping portion of the character pattern without separating the character pattern and the non-character pattern in the input image, and the character contacting the character pattern. The character pattern can be recognized regardless of the shape of the pattern other than the above and the manner of contact. In addition, since the image is divided into meshes, the similarity is obtained for each mesh, and the overall similarity is obtained from the consistency of those positions, and recognition is performed, so recognition is possible without having to cut out character areas from the image in character units. become.

  In addition, a present Example is an example to the last and does not limit this invention. The present invention can be implemented by appropriately changing the configuration and operation within the scope of the technical idea described in the claims.

  For example, in this embodiment, the character recognition device 1 is configured to perform switching between hash table registration by the hash table registration unit 22 and character recognition by the recognition processing unit 21. You may comprise so that only the character recognition may be performed using the created hash table.

  In addition, an arbitrary method can be applied to the calculation of the characteristics of the partial input image and the partial character image, the search for the character category having similar characteristics, and the consistency determination of the positional relationship between the partial input image and the partial character image. . For example, in this embodiment, when the path between the nodes is drawn, the consistency of the positional relationship is determined based on the relative direction between the meshes. However, the positional relationship is also matched with respect to the distance between the meshes in addition to the relative direction. It can also be used as a criterion for sex.

  The character recognition method according to the present invention can also be implemented by realizing each functional unit with dedicated hardware as in the character recognition device 1 shown in FIGS. 1 and 2. It can also be implemented by causing the hardware resources of the computer to operate as respective functional units.

  FIG. 18 is a schematic configuration diagram illustrating a schematic configuration of a computer that executes a character recognition program. As shown in the figure, the computer 40 includes an input device 41, a display device 42, an image reading device 43, an external interface 44, a CPU (Central Processing Unit) 45, a ROM (Read Only Memory) 46, a RAM (Random Access Memory). ) 47 and HDD (Hard Disk Drive) 48 are connected by a bus 49. Here, the input device 41 corresponds to the input unit 11 shown in FIG. 1, the display device 42 also corresponds to the display unit 12, and the image reading device 43 and the external interface 44 also correspond to the image reading unit 13 and the external interface 14. Corresponding to

  The ROM 46 stores in advance a recognition processing program 51 and a hash table registration program 52 that exhibit the same functions as the recognition processing unit 21 and the hash table registration unit 22 described above.

  The CPU 45 can implement its functions by reading the recognition processing program 51 and the hash table registration program 52 from the ROM 46 and executing them. For example, in the example shown in FIG. 18, the CPU 45 reads the recognition processing program 51 from the ROM 46 and operates the recognition processing process 54, and the CPU 45 performs the same operation as the recognition processing unit 21.

  The HDD 48 stores hash table data 53 as shown in FIG. The hash table data 53 is read by the CPU 45 and expanded on the RAM 47, and operates as the hash table 16 shown in FIGS.

  By the way, the above-described recognition processing program 51 and hash table registration program 52 are not necessarily stored in the ROM 46 from the beginning. For example, a flexible disk (FD), a CD-ROM, an MO disk, “General physical media” such as DVD discs, magneto-optical discs, IC cards, etc., or “fixed physical media” such as hard disk drives (HDDs) provided inside and outside the computer 40, as well as public lines, the Internet Each program may be stored in “another computer (or server)” connected to the computer 40 via a LAN, WAN, etc., and the computer 40 may read and execute each program from these. .

(Supplementary note 1) A character recognition program for causing a computer to execute a character recognition process for recognizing characters in an input image,
A dividing step of dividing the input image into a plurality of partial input images;
A feature amount calculating step for calculating a feature amount for each of the plurality of partial input images obtained by the dividing step;
Characteristic quantities of partial character images in each part obtained by dividing a plurality of character images into a plurality of parts, the positional relationship of each partial character image in the character image, and the character type indicated by the character image A search step for searching for a partial character image having a similar feature value for each of the partial input images, using the feature value calculated by the feature value calculation step as a key,
A plurality of partial character images obtained by the search step are aggregated for each character type, the positional relationship of the partial character images in each character type, and the partial input image corresponding to the partial character image in the input image A consistency determining step of determining consistency with the positional relationship and extracting a combination of partial input images that matches the positional relationship of the partial character images as character candidates;
When the character candidates extracted by the consistency determination step have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as constituent elements of the character indicated by the character type. A recognition step;
A character recognition program for causing a computer to execute.

(Additional remark 2) The said consistency determination step makes a partial input image similar to the partial character image of the same character classification among the said some partial input images a node, and the relative position between the partial character images corresponding to each node A graph creation step of creating a graph by connecting nodes in which the relationship and the relative positional relationship between the partial input images corresponding to each node are consistent, and a complete subgraph of the graph created by the graph creation step The character recognition program according to claim 1, further comprising a clique extraction step of extracting a clique as the character candidate.

(Appendix 3) In the recognition step, when the character candidates extracted in the consistency determination step have a predetermined number or more of partial input images, a rectangle circumscribing the combination of the partial input images is obtained, and the rectangle is included in the rectangle. The character recognition program according to appendix 1 or 2, characterized in that the character indicated by the character type is recognized.

(Additional remark 4) The said search table is the feature-value of each partial character image with respect to the partial character image obtained when the division processing from which division size and / or a division position differ with respect to the same character image, respectively. The character recognition program according to any one of appendices 1 to 3, wherein a positional relationship between the partial character images in the character image and a character type indicated by the character image are stored in association with each other. .

(Supplementary Note 5) The search table associates a plurality of similar feature amounts with the partial character image, and the positional relationship of each partial character image in the character image with respect to each of the plurality of feature amounts. The character recognition program according to any one of appendices 1 to 4, wherein the character type indicated by the character image is stored in association with each other.

(Appendix 6) A character recognition device for recognizing characters in an input image,
Characteristic quantities of partial character images in each part obtained by dividing a plurality of character images into a plurality of parts, the positional relationship of each partial character image in the character image, and the character type indicated by the character image Search table storage means for storing a search table associated with
Dividing means for dividing the input image into a plurality of partial input images;
Feature quantity calculating means for calculating a feature quantity for each of the plurality of partial input images obtained by the dividing means;
Search means for searching for partial character images having similar feature quantities for each of the partial input images with reference to the search table using the feature quantities calculated by the feature quantity calculating means as a key;
A plurality of partial character images obtained by the search means are aggregated for each character type, the positional relationship of the partial character images in each character type, and the partial input image corresponding to the partial character image in the input image Consistency determining means for determining consistency with the positional relationship and extracting a combination of partial input images that matches the positional relationship of the partial character images as character candidates;
When the character candidates extracted by the consistency determination unit have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as the constituent elements of the character indicated by the character type. Recognition means;
A character recognition device comprising:

(Additional remark 7) The said consistency determination means uses as a node the partial input image similar to the partial character image of the same character type among these partial input images, and the relative position between the partial character images corresponding to each node Connecting the nodes where the relationship and the relative positional relationship between the partial input images corresponding to each node do not contradict each other, creating a graph, and extracting a clique that is a complete subgraph of the created graph as the character candidate The character recognition device according to appendix 6, which is characterized.

(Supplementary Note 8) When the character candidates extracted by the consistency determination unit have a predetermined number or more of partial input images, the recognition unit obtains a rectangle circumscribing the combination of the partial input images, and includes the rectangle within the rectangle. The character recognition device according to appendix 6 or 7, wherein the character identified by the character type is recognized to exist.

(Additional remark 9) The said search table is the feature-value of each partial character image with respect to the partial character image obtained when the division processing from which division size and / or a division position differ with respect to the same character image, respectively. The character recognition device according to any one of appendices 6 to 8, wherein a positional relationship between the partial character images in the character image and a character type indicated by the character image are stored in association with each other. .

(Supplementary Note 10) The search table associates a plurality of similar feature amounts with the partial character image, and a positional relationship of each partial character image in the character image with respect to each of the plurality of feature amounts; The character recognition device according to any one of appendices 6 to 9, wherein the character type indicated by the character image is stored in association with each other.

(Supplementary note 11) A character recognition method for recognizing characters in an input image,
A dividing step of dividing the input image into a plurality of partial input images;
A feature amount calculating step for calculating a feature amount for each of the plurality of partial input images obtained by the dividing step;
Characteristic quantities of partial character images in each part obtained by dividing a plurality of character images into a plurality of parts, the positional relationship of each partial character image in the character image, and the character type indicated by the character image A search step for searching for a partial character image having a similar feature amount for each of the partial input images, using the feature amount calculated by the feature amount calculation step as a key,
A plurality of partial character images obtained by the search step are aggregated for each character type, the positional relationship of the partial character images in each character type, and the partial input image corresponding to the partial character image in the input image A consistency determining step of determining consistency with the positional relationship and extracting a combination of partial input images that matches the positional relationship of the partial character images as character candidates;
When the character candidates extracted by the consistency determination step have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as the constituent elements of the character indicated by the character type. Recognition process;
A character recognition method characterized by comprising

(Additional remark 12) The said consistency determination process makes a partial input image similar to the partial character image of the same character classification among a plurality of said partial input images a node, and the relative position between the partial character images corresponding to each node A graph creation step of creating a graph by connecting nodes in which the relationship and the relative positional relationship between partial input images corresponding to each node do not contradict each other, and a complete subgraph of the graph created by the graph creation step The character recognition method according to claim 11, further comprising a clique extraction step of extracting a clique as the character candidate.

(Additional remark 13) The said recognition process calculates | requires the rectangle which circumscribes the combination of the said partial input image, when the character candidate extracted by the said consistency determination process has a predetermined number or more of partial input images, and in the said rectangle 13. The character recognition method according to appendix 11 or 12, wherein a character indicated by the character type is recognized.

(Additional remark 14) The said search table is the feature-value of each partial character image with respect to the partial character image obtained when the division processing from which division size and / or a division position differ with respect to the same character image, respectively. The character recognition method according to any one of appendices 11 to 13, wherein a positional relationship between the partial character images in the character image and a character type indicated by the character image are stored in association with each other. .

(Supplementary Note 15) The search table associates a plurality of similar feature amounts with the partial character image, and the positional relationship of each partial character image in the character image with respect to each of the plurality of feature amounts. The character recognition method according to any one of appendices 11 to 14, wherein the character type indicated by the character image is stored in association with each other.

  As described above, the character recognition program, the character recognition device, and the character recognition method according to the present invention are useful for character recognition, and are particularly suitable for recognition of character images in which patterns other than characters are superimposed.

It is a functional block diagram which shows the functional structure of the character recognition apparatus concerning the Example of this invention. 4 is a functional configuration diagram illustrating the functional configuration of a recognition processing unit 21 in more detail. It is explanatory drawing explaining the registration of the hash table 16 by the hash table registration part 22. FIG. It is explanatory drawing explaining the feature calculation and character category search of the input image by the recognition process part. It is explanatory drawing explaining the total for every character category of a search result. It is explanatory drawing explaining the graphing process of the partial input image by the position consistency determination part. It is explanatory drawing explaining the path connection between nodes. It is explanatory drawing explaining extraction and evaluation of a clique. It is explanatory drawing explaining the recognition of the character area by the character determination part. It is a figure which shows the voting result data which the position consistency determination part 35 produces from the search result by a character category search part. It is a figure which shows the graph data which the position consistency determination part 35 produces. It is a flowchart explaining the processing operation | movement of hash table registration. It is a flowchart explaining the processing operation | movement in the recognition process of a character. It is explanatory drawing explaining the modification of normalization at the time of hash table creation. It is explanatory drawing explaining the case where a several different mesh division | segmentation is performed with respect to one character image. It is explanatory drawing explaining the recognition process in the case of using a (n, dx, dy) feature. 1 is a schematic configuration diagram illustrating a schematic configuration of an image recognition system using the present invention. It is a schematic block diagram explaining the schematic structure of the computer which performs a character recognition program.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Character recognition apparatus 11 Input part 12 Display part 13 Image reading part 14 External interface 15 Memory | storage part 16 Hash table 20 Control part 21 Recognition process part 22 Hash table registration part 31 Mesh division part 32 Feature calculation part 33 Normalization process part 34 Character Category search unit 35 Position consistency determination unit 36 Character determination unit 40 Computer 41 Input device 42 Display device 43 Image reader 44 External interface 49 Bus 51 Recognition processing program 52 Hash table registration program 53 Hash table data 54 Recognition processing process 100 Image recognition System 101 Scanner 102 Difference image creation unit 103 Image analysis unit D1, D2, D3 Image data D4 Analysis result data G1 Graph G2, G3 Clique m21, m43, m44, m104, m 05, m108, mα, mβ mesh mα ', mα'',mβ' projected image

Claims (7)

A character recognition program for causing a computer to execute character recognition processing for recognizing characters in an input image,
A division step of dividing the input image into a plurality of partial input images having a size in which one character in the input image is divided into a predetermined number ;
A feature amount calculating step for calculating a feature amount for each of the plurality of partial input images obtained by the dividing step;
A character quantity of a partial character image obtained by dividing a plurality of character images into the predetermined number , a character image of which character image has a feature quantity of the partial character image, and a character quantity of the partial character image at which position. A partial character image having a similar feature amount for each of the partial input images is searched from the search table in which the character type indicated by is stored in association with each other using the feature amount calculated in the feature amount calculating step as a key. A search step;
A plurality of partial character images obtained by the search step are aggregated for each character type, the positional relationship of the partial character images in each character type, and the partial input image corresponding to the partial character image in the input image A consistency determining step of determining consistency with the positional relationship and extracting a combination of partial input images that matches the positional relationship of the partial character images as character candidates;
When the character candidates extracted by the consistency determination step have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as constituent elements of the character indicated by the character type. A recognition step;
A character recognition program for causing a computer to execute.   The consistency determination step uses a partial input image similar to a partial character image of the same character type as a node among the plurality of partial input images, and a relative positional relationship between partial character images corresponding to each node, A graph creation step of creating a graph by connecting nodes in which the relative positional relationship between the partial input images corresponding to the nodes is consistent, and a clique that is a complete subgraph of the graph created by the graph creation step The character recognition program according to claim 1, further comprising a clique extraction step of extracting as a candidate.   In the recognition step, when the character candidates extracted in the consistency determination step have a predetermined number or more of partial input images, a rectangle circumscribing the combination of the partial input images is obtained, and the rectangle includes the rectangle according to the character type. The character recognition program according to claim 1, wherein the character recognition unit recognizes that the indicated character exists. The lookup table, for the same partition size and / or split position different division respectively obtained when the processing was performed partial character image to a character image, and the feature quantity of each partial character images, the respective parts 4. The character image of which character image is a character image at which position of a character image and a character type indicated by the character image are stored in association with each other. The character recognition program as described in any one. The search table associates a plurality of similar feature quantities with the partial character image , and for each of the plurality of feature quantities, a part of which character image has a feature quantity of the partial character image. The character recognition program according to any one of claims 1 to 4, wherein a character image feature value and a character type indicated by the character image are stored in association with each other. A character recognition device for recognizing characters in an input image,
A dividing unit that divides the input image into a plurality of partial input images having a size in which one character in the input image is divided into a predetermined number ;
Feature quantity calculating means for calculating a feature quantity for each of the plurality of partial input images obtained by the dividing means;
A character quantity of a partial character image obtained by dividing a plurality of character images into the predetermined number , a character image of which character image has a feature quantity of the partial character image, and a character quantity of the partial character image at which position. Search table storage means for storing a search table in association with the character type indicated by
Search means for searching for partial character images having similar feature quantities for each of the partial input images with reference to the search table using the feature quantities calculated by the feature quantity calculating means as a key;
A plurality of partial character images obtained by the search means are aggregated for each character type, the positional relationship of the partial character images in each character type, and the partial input image corresponding to the partial character image in the input image Consistency determining means for determining consistency with the positional relationship and extracting a combination of partial input images that matches the positional relationship of the partial character images as character candidates;
When the character candidates extracted by the consistency determination unit have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as the constituent elements of the character indicated by the character type. Recognition means;
A character recognition device comprising: A character recognition method for recognizing characters in an input image,
A division step of dividing the input image into a plurality of partial input images having a size in which one character in the input image is divided into a predetermined number ;
A feature amount calculating step for calculating a feature amount for each of the plurality of partial input images obtained by the dividing step;
A character quantity of a partial character image obtained by dividing a plurality of character images into the predetermined number , a character image of which character image has a feature quantity of the partial character image, and a character quantity of the partial character image at which position. A partial character image having a similar feature value for each of the partial input images is searched from the search table in which the character type indicated by is stored in association with each other using the feature value calculated by the feature value calculation step as a key. Search process;
A plurality of partial character images obtained by the search step are aggregated for each character type, the positional relationship of the partial character images in each character type, and the partial input image corresponding to the partial character image in the input image A consistency determining step of determining consistency with the positional relationship and extracting a combination of partial input images that matches the positional relationship of the partial character images as character candidates;
When the character candidates extracted by the consistency determination step have a predetermined number or more of partial input images, the partial input images of the character candidates are recognized as the constituent elements of the character indicated by the character type. Recognition process;
A character recognition method characterized by comprising

Images